Copilot computer use or Sume's MCP for image and video?

Use Sume's hosted MCP server (or its API) for image and video generation from Copilot, and keep computer use for software that has no API, command line or MCP integration. Computer use clicks through a desktop app; Sume's tools return a job you can wait on, read and bill-check.

GitHub described computer use on 1 October 2026 in GitHub Copilot can now interact with desktop apps with computer use (read 2026-10-02). It is in public preview in Copilot CLI and the Copilot app on macOS and Windows. The Sume side is from the MCP overview and MCP tools and gates.

What does Copilot computer use do?

Per the changelog, Copilot can read accessible app content and visual context, click controls, enter and edit text, press keys, scroll, drag and navigate workflows across applications. GitHub frames it as a way to automate legacy and GUI-only software without an API, command-line interface or MCP integration.

Control is approval based. Copilot asks before controlling an app, and you can review or reset the apps you chose to always allow. In the CLI the switch is /computer on, /computer show and /computer off; in the Copilot app it is Settings, Computer Use. On macOS you grant Accessibility and Screen Recording permissions, and organization-managed settings can turn it off.

Where does Sume sit instead?

Sume exposes a hosted MCP endpoint at https://mcp.sume.com/mcp with tools for account, catalog, jobs, assets, generation, crawl and Avatar work. Router stills and clips are generate_image and generate_video; music, speech and transcription are music_create, tts_create and stt_create. A client that has the server configured calls these directly, with no screen involved.

That difference shows up in what you can check afterwards. A generation call returns a job id; jobs_wait and jobs_result tell you whether it finished, and usage_get and balance_get show spend. A click sequence in a desktop app leaves you with whatever the app displays.

How do the two compare for a generation task?

The table summarizes the trade as the vendor pages and Sume's docs describe it. It makes no speed claim, because neither page publishes one.

How do I check that the Sume side works before trusting an agent with it?

Point the client at https://mcp.sume.com/mcp, finish the OAuth sign-in on the MCP host, and leave Write off. Then ask the agent to call mcp_health, which confirms the endpoint, the auth source and the safety posture, followed by tools_list, which lists every tool visible to that session. With mcp:read only you should see read-only tools, and any attempt at a paid call returns insufficient_scope. That is a useful dry run for any agent that can also click through your desktop, because it proves the agent can read your Sume account state without being able to spend.

When you do want generation, enable Write on the consent page, call tools_schema for the tool you plan to use, and run it once with dry_run: true. The docs are explicit that idempotency_key is required on write and paid tools and is not a human approval, so keep your own confirmation step in the workflow. For a deeper comparison of surfaces, see MCP vs CLI vs API for AI agents.

When would I still turn computer use on?

Turn it on when the last step lives in an app that Sume cannot reach: a desktop editor, a broadcast tool, a client portal with no API. A reasonable split is that Sume generates and stores the asset, and computer use drops the finished file into the GUI-only tool. Sume's assets_download_url gives the asset link; the click-through part is Copilot's.

Do not use computer use to drive the Sume web app to submit paid jobs. The MCP route has the idempotency_key and dry_run gates and the OAuth scope split; a mouse-driven submit has none of them. Also note that Sume's mcp:read OAuth sessions never see paid tools, so a read-only connection is a safe default while you try Copilot out. Sume documents a generic remote MCP connection rather than a Copilot-specific integration, so the Copilot clients connect to the server like any other HTTP MCP client.